Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley

Frank Coyle emphasizes the importance of active creation and experimentation over passive reading, drawing inspiration from Sister Corita Kent and John Cage. He advocates for a philosophy where mistakes are not failures but opportunities to learn through 'making,' encouraging students to engage in hands-on activities like writing and drawing rather than solely relying on typing.

Coyle defines ontologies as 'a formal specification of a shared conceptualization,' emphasizing their role in providing agents with structured knowledge about the world. He explains that this concept, initially explored by Gruber in 1993, is crucial for representing entities, their properties, and relationships within a domain, enabling more effective agent reasoning.

Coyle highlights the significant trend of neuro-symbolic AI, which involves integrating probabilistic agents (like LLMs) with formal representations like ontologies. This convergence aims to leverage the creativity of LLMs while grounding them in structured knowledge and mitigating issues such as hallucinations that arise from purely probabilistic models.

Coyle contrasts graph databases with relational databases to explain the rise of ontologies. He notes that traditional relational databases, which store data in tables, are restrictive and require significant restructuring when new information is added. Graph databases offer more flexibility by allowing for easy attachment of properties and relationships, making them well-suited for representing complex knowledge structures.

The speaker explains that Large Language Models (LLMs) fundamentally operate by predicting the next word with a given probability, and therefore cannot inherently 'do' anything. To enable them to interact with external systems, they are used as generators of parameters for tools. The LLM is prompted to determine how a tool can help achieve a desired outcome, then sets up the necessary parameters and returns those to an external executor.

Following tool execution, ontologies are introduced as a mechanism for validating the LLM's output. The speaker emphasizes that after a tool is called and provides information, this data needs to be formatted in a way that can be understood by a validator operating within an ontological framework. This allows for reasoning about whether the response from the LLM aligns with established domain knowledge.

To address the lack of type safety in Python, Pydantic is presented as a solution. Pydantic allows developers to specify and enforce data types for parameters, preventing unexpected behavior and improving code reliability. The speaker recommends using Pydantic to check parameter types and then employing ontologies to validate the overall results.

The speaker introduces OWL (Web Ontology Language) as a tool for creating logical constructs that can detect errors and provide guardrails for LLMs. Examples are given, such as identifying duplicate refunds or ensuring payout destinations are correct, demonstrating how ontologies can catch inconsistencies that might be difficult to identify using natural language alone.